[2608.00065] H+ Embedding: Harmonizing World and Token-Stage Retrieval with Context-Dependent Phrases

View a PDF of the paper titled H+ Embedding: Harmonizing World and Token-Stage Retrieval with Context-Dependent Phrases, by Shusen Zhang and eight different authors

View PDF
HTML (experimental)

Summary:Terminology-intensive retrieval, particularly in medical settings, is determined by preserving multi-word entities, abbreviations, numerical constraints, and compositional ideas. Nevertheless, present representations lie at two extremes: single-vector retrievers typically over-compress native relevance indicators, whereas token-level late interplay retains each tokenizer subword at substantial indexing, storage, and scoring price. This mismatch raises a pure query: can context-dependent phrases present a helpful retrieval unit between world vectors and tokens? We introduce H+ Embedding, a unified multi-granularity retriever that predicts variable-length phrase partitions, preserves uncovered tokens as singletons, and applies importance-guided unit choice with weighted MaxSim interplay. Throughout 16 scientific, medical, and bilingual duties, its phrase retrieval department exceeds the worldwide retrieval department by 6.91 macro nDCG@10. It additionally practically matches Token whereas utilizing 13.7% fewer doc vectors and outperforms content-independent grouping guidelines underneath average vector budgets. Context-dependent phrase interplay due to this fact supplies an intermediate quality-cost level between world compression and token-level interplay for sensible retrieval methods.